• GeForce2 announcement

    From mohd.tahauddin@gmail.com@21:1/5 to Tom Duff on Tue Feb 23 02:58:05 2016
    Hello Tom:

    I am not even sure if you are still following this :)

    I read about your original rebuttal to NVIDIA's marketing hype 16 years ago and for some reason, I wanted to hear your thoughts about modern 3D accelerators with programmable shading languages.

    How would you compare a modern accelerator in the following context:
    1) Replacing/complimenting a render farm from 16 years ago. Do you think a modern graphics card begins to approach the original promise from NVIDIA (comparing against a render farm from 2000).
    2) A modern graphics card vs a modern render farm

    I would really love to hear your thoughts, especially, what needs to change/improve, how much more complex shaders in offline renderers are vs a modern graphics card etc.


    On Monday, May 1, 2000 at 12:00:00 AM UTC-7, Tom Duff wrote:
    Allen McPherson wrote:
    I'm curious. Obviously, real-time TS2 frame generation is
    very difficult, especially given the required data rates you
    provided. On the other hand, would it be useful to use this
    technology to preview shaders, different animation scenarios,
    develop new shading algorithms, etc? [on lower resolution
    models and image resolutions of course]

    We certainly use 3D accelerators (SGI rather than NVIDIA) to preview animation. Until you can compile shading language programs to run on them, they won't be much use for shading and lighting.

    Also, rather than one on every desk, what about one in each of
    your 1000+ nodes of the render farm? We work in very different
    domains, but we're looking at building just such a system (though
    only on 32-64 nodes for now).

    3D accelerators mostly don't do anyting we're interested in doing a lot of. Of the 1.2 million hours of CPU time that goes into making a set of TS2 frames, about 1.1 million hours is devoted to executing shading language code, for which NVIDIA's cards are essentially no help at all. If they could cache texture off a several-gigabyte UNIX filesystem and pull filtered texture samples at arbitrary coordinates out at the rates they advertise, they might be some use, since I think about half of the time we spend in shading is devoted to sampling texture maps. But note that by Ahmdahl's
    law, if their boards were infinitely fast, we'd stilly only see a 50% rendering speed-up.

    Tom Duff. Some sort of background check is in order.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)